A focused AI automation costs $6,000 to $20,000 to build, plus $50 to $500 a month to run. That is the short answer. The longer answer is that the build price is the part everyone quotes and the part that matters least, because what determines whether the project pays back is the running cost, the failure handling, and whether you needed AI at all.
What it costs to build
| Scope | Typical build cost | What it covers | Right for |
|---|---|---|---|
| Single task | $2,000–6,000 | One narrow job — classify inbound email, extract fields from a form | Testing whether this works for you |
| One workflow | $6,000–20,000 | Intake through to routing and drafted response, wired into your systems | Most small businesses starting out |
| Multi-process | $20,000–50,000 | Several connected workflows, shared data, reporting and monitoring | Established operations with real volume |
| Embedded in software | $40,000+ | AI features built into a custom platform as one component of a system | Product businesses, not internal admin |
Southern Ontario market rates, 2026. Build cost only — model usage is billed separately and is covered below.
The running cost nobody puts in the quote
This is the real difference from ordinary automation, which costs nothing to run once built. Every task a model handles has a usage cost, billed by volume of text processed. At small business scale that lands somewhere between $50 and $500 a month.
That is almost always trivial next to the labour it replaces. It is also a real, recurring line item, and it belongs in the numbers before you sign rather than arriving as a surprise in month two. If a quote does not mention it, the person quoting either has not built one of these before or is hoping you will not ask.
Working out the payback
The arithmetic is straightforward and worth doing yourself before anyone sells you anything. Take the hours the task consumes each week, multiply by a realistic loaded hourly rate for whoever does it, and annualise. Compare that against build cost plus twelve months of running cost.
A worked example
A trades business spends ten hours a week on quote intake and follow-up. At a loaded rate of $35 an hour that is $18,200 a year. A $15,000 build with $150 a month in usage costs $16,800 in year one and $1,800 a year after that. It pays back in about eleven months and then keeps paying. If your equivalent number is three hours a week, it does not, and you should not buy it.
Where the money actually goes
Not on the model. Getting a demo working takes an afternoon, which is why the demos are so convincing and so misleading. The cost sits in everything after: connecting to the systems you already run, deciding what happens when the model is wrong, keeping a person in the loop where being wrong is expensive, and testing against real messy inputs from your business rather than the tidy examples that always work.
- ✓Integration with your existing systems — usually the single largest line
- ✓Failure handling and escalation paths for low-confidence cases
- ✓Testing against your real historical inputs, not sample data
- ✓Monitoring, so failures surface on a dashboard rather than via a customer
When the honest answer is not to buy it
Often. If your bottleneck is that two systems do not talk to each other, that is an integration and it costs less. If nobody follows up on quotes, that is a workflow problem and a rule fixes it. If the task happens twice a week, the payback maths will not work no matter how good the technology is.
We scope the rules-based version alongside the AI one specifically so you can see both numbers. A fair number of the businesses we talk to walk away with the cheaper option, which is the correct outcome.